π― What it does: The ALIDIFF framework is proposed to align pre-trained target-aware molecular diffusion models to enhance the binding energy and structural rationality between ligands and targets.
Alignment at Pre-training! Towards Native Alignment for Arabic LLMs
Juhao Liang (Shenzhen Research Institute of Big Data), Jinchao Xu (King Abdullah University of Science and Technology)
CodeTransformerLarge Language ModelText
π― What it does: Proposed and implemented a 'Native Alignment' method during the pre-training phase, focusing on Arabic LLM, and released two open-source models: LLaMA3-Tamed-8B and LLaMA3-Tamed-70B.
Yuqing Yang (Fudan University), Pengfei Liu (Shanghai Jiao Tong University)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: A framework for aligning the honesty of large language models is proposed, encouraging models to proactively refuse to answer when they do not know the answer.
π― What it does: A multi-path aggregation (MPA) model is proposed to achieve an all-in-one unified architecture for human-machine vision joint image coding.
Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization
Xinyu Lyu (Southwestern University of Finance and Economics), Jingkuan Song (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China)
CodeRetrievalOptimizationTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: A strategy called Hallucination-Induced Optimization (HIO) is proposed to enhance contrastive decoding by inducing more hallucinations, significantly reducing the hallucination output of large visual language models.
Almost Minimax Optimal Best Arm Identification in Piecewise Stationary Linear Bandits
Yunlong Hou (National University of Singapore), Zixin Zhong (Hong Kong University of Science and Technology)
CodeReinforcement LearningTime Series
π― What it does: A piecewise stationary linear bandit (PSLB) model is proposed, and the PS Ξ΅ BAI+ algorithm is designed to identify the Ξ΅-optimal arm under a given confidence level Ξ΄.
Almost Surely Asymptotically Constant Graph Neural Networks
Sam Adam-Day (University of Oxford), Ben Finkelshtein (University of Oxford)
CodeClassificationGraph Neural NetworkGraph
π― What it does: The convergence of the output of real-valued GNN classifiers with respect to graph size is studied under random graph models, proving that most GNNs ultimately output a constant.
Almost-Linear RNNs Yield Highly Interpretable Symbolic Codes in Dynamical Systems Reconstruction
Manuel Brenner (Heidelberg University), Daniel Durstewitz (Heidelberg University)
CodeExplainability and InterpretabilityRecurrent Neural NetworkTime SeriesBiomedical DataElectrocardiogram
π― What it does: Proposes and trains an Almost Linear Recursive Neural Network (AL-RNN) that automatically learns the simplest piecewise linear representation from time series data and generates interpretable symbolic encodings.
AlphaMath Almost Zero: Process Supervision without Process
Guoxin Chen (Tongyi Lab), Kai Fan (Tongyi Lab)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: AlphaMath automatically generates high-quality mathematical reasoning processes by combining a pre-trained LLM with Monte Carlo Tree Search (MCTS) and a lightweight value model, without the need for manual or GPT-4 process annotations.
AlphaPruning: Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models
Haiquan Lu (Nankai University), Yaoqing Yang (University of California)
CodeTransformerLarge Language ModelText
π― What it does: This paper proposes a hierarchical sparsification method called AlphaPruning based on the theory of reparameterization self-regularization (HT-SR), using the ESD shape index (PL Alpha Hill) to allocate the hierarchical sparsity ratio of LLM layers, further enhancing the sparsity rate while maintaining performance.
ALPS: Improved Optimization for Highly Sparse One-Shot Pruning for Large Language Models
Xiang Meng (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
CodeOptimizationLarge Language ModelText
π― What it does: An optimized single-shot sparse pruning framework called ALPS is proposed, which uses ADMM to solve non-convex β0 constraints, combined with dynamic penalty parameter updates and PCG post-processing, achieving high-quality one-click pruning for large-scale LLMs.
π― What it does: A reinforcement learning-based pixel attack method called RFPAR, which utilizes memory and forgetting mechanisms, is proposed to achieve black-box attacks with minimal pixel perturbation in image classification and object detection.
AmoebaLLM: Constructing Any-Shape Large Language Models for Efficient and Instant Deployment
Yonggan Fu (Georgia Institute of Technology), Yingyan Celine Lin (Georgia Institute of Technology)
CodeCompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The AmoebaLLM framework is proposed, which can instantly extract sub-networks of any depth/width combination after a one-time fine-tuning, allowing large language models to be flexibly and efficiently deployed across various platforms and application scenarios.
CodeDomain AdaptationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextBiomedical Data
π― What it does: This paper proposes an adaptable modular knowledge agent (AMOR) based on finite state machines (FSM), capable of reasoning, retrieving, and answering from knowledge bases in different domains through process feedback.
Amortized Planning with Large-Scale Transformers: A Case Study on Chess
Anian Ruoss (Google DeepMind), Tim Genewein (Google DeepMind)
CodeReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningTabularBenchmark
π― What it does: Using a large-scale Transformer to predict action values on a chessboard, and based on this, constructing a search-free game strategy, evaluating its playing strength against traditional search engines.
An Autoencoder-Like Nonnegative Matrix Co-Factorization for Improved Student Cognitive Modeling
Shenbao Yu (Fujian Normal University), Mingwei Lin (Fujian Normal University)
CodeOptimizationExplainability and InterpretabilityAuto EncoderTabular
π― What it does: A non-negative matrix co-factorization model based on autoencoders (AE-NMCF) is proposed to simultaneously predict students' homework scores and estimate their mastery of knowledge concepts.
An eye for an ear: zero-shot audio description leveraging an image captioner with audio-visual token distribution matching
Hugo Malard (Telecom Paris), Slim Essid (Telecom Paris)
CodeGenerationData SynthesisRepresentation LearningTransformerPrompt EngineeringVision Language ModelVideoMultimodalityAudio
π― What it does: This paper proposes an unsupervised zero-shot audio description method that utilizes an image description model to map audio features into a visual encoding space through distribution alignment, thereby achieving natural language descriptions of audio.
ANAH-v2: Scaling Analytical Hallucination Annotation of Large Language Models
Yuzhe Gu (Shanghai Jiao Tong University), Kai Chen (Shanghai AI Laboratory)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
π― What it does: This paper proposes an EM-based iterative self-training framework that gradually expands the hallucination labeled dataset and improves annotator accuracy.
Animal-Bench: Benchmarking Multimodal Video Models for Animal-centric Video Understanding
Yinuo Jing (Beijing University of Posts and Telecommunications), Jun Guo (Beijing University of Posts and Telecommunications)
CodeLarge Language ModelDiffusion modelVideoMultimodalityBenchmark
π― What it does: This study constructed the animal-centered multimodal video evaluation benchmark Animal-Bench and evaluated the performance of several large models based on it.
π― What it does: Proposes Annealed Multiple Choice Learning (aMCL), which combines deterministic annealing with MCL training to address hypothesis collapse and local optima issues.
ANT: Adaptive Noise Schedule for Time Series Diffusion Models
Seunghan Lee (Yonsei University), Taeyoung Park (Yonsei University)
CodeDiffusion modelTime Series
π― What it does: An adaptive noise scheduling method based on the non-stationarity statistics of time series (ANT) is proposed, which can automatically select appropriate noise schedules for each dataset before training the diffusion model;
π― What it does: Proposes the Any2Graph framework, which supports arbitrary inputs (images, text, vectors, etc.) to directly predict graph structures with arbitrary size and unordered nodes through an end-to-end neural network.
Apathetic or Empathetic? Evaluating LLMs' Emotional Alignments with Humans
Jen-tse Huang (Chinese University of Hong Kong), Michael Lyu
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: By constructing the EmotionBench framework, experiments were conducted on 428 emotion-inducing scenarios to measure the changes in emotional states of LLMs when faced with negative emotional situations, and these results were compared with the questionnaire responses of 1,266 human subjects.
π― What it does: This paper proposes to limit the Classifier-Free Guidance (CFG) during the sampling process of diffusion models to a specific range of noise levels, using guidance only in the middle noise range, thereby improving generation quality and speed.
Approximately Pareto-optimal Solutions for Bi-Objective k-Clustering
Anna Arutyunova (Heinrich Heine University DΓΌsseldorf), Julian Wargalla (Heinrich Heine University DΓΌsseldorf)
CodeOptimizationTabularTime Series
π― What it does: This paper studies the bi-objective clustering problem, proposing algorithms to approximate the Pareto front and exploring combinations of multi-objective clustering based on metrics such as k-center, k-diameter, k-median, k-means, and k-separation.
CodeRepresentation LearningDrug DiscoveryAuto EncoderTabularBiomedical Data
π― What it does: Proposed and implemented the LMI (latent mutual information) approximation method, which uses low-dimensional learning representations and employs a non-parametric MI estimator in this space, addressing the challenge of MI estimation for high-dimensional variables.
Natalie Maus (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
CodeOptimizationDrug DiscoveryTabular
π― What it does: This paper proposes a joint optimization of the posterior approximation of Sparse Variational Gaussian Processes (SVGP) and the sampling decisions of Bayesian Optimization (BO), forming the Expected Utility Lower Bound (EULBO) to enhance the data acquisition efficiency of high-dimensional large-budget BO.
AR-Pro: Counterfactual Explanations for Anomaly Repair with Formal Properties
Xiayan Ji (University of Pennsylvania), Insup Lee (University of Pennsylvania)
CodeAnomaly DetectionExplainability and InterpretabilityDiffusion modelImageTime Series
π― What it does: Proposes the AR-Pro framework, which uses linearly decomposable anomaly detectors to generate counterfactuals for anomalous samples, serving as an explanation for interpretability.
ARC: A Generalist Graph Anomaly Detector with In-Context Learning
Yixin Liu (Griffith University), Shirui Pan (Hong Kong Polytechnic University)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: A 'one-for-all' graph anomaly detection model ARC is proposed, which can instantly detect anomalous nodes in the graph using a small amount of contextual information from normal nodes without retraining or fine-tuning the target dataset.
Are Graph Neural Networks Optimal Approximation Algorithms?
Morris Yau (Massachusetts Institute of Technology), Stefanie Jegelka (Technical University of Munich and Massachusetts Institute of Technology)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: This paper designs and implements a graph neural network architecture named OptGNN, which can capture information from optimal approximation algorithms (based on semidefinite programming) and learn high-quality approximate solutions through unsupervised training.
Are High-Degree Representations Really Unnecessary in Equivariant Graph Neural Networks?
Jiacheng Cen (Renmin University of China), Wenbing Huang (Renmin University of China)
CodeGraph Neural NetworkGraph
π― What it does: This paper challenges the assumption of the necessity of higher-order transferable vectors in equivariant graph neural networks, proposing the HEGNN model, which enhances expressive power by introducing higher-order transferable vectors while maintaining the efficiency of EGNN.
π― What it does: This paper proposes a method called LPLD (Label Pruning for Large-scale Distillation), which first batches similar samples during the image synthesis phase and incorporates class-level BN supervision to enhance the intra-class diversity of the synthesized images. Subsequently, it randomly crops the soft labels and resamples from an improved label pool, achieving soft label compression and performance enhancement.
Are More LLM Calls All You Need? Towards the Scaling Properties of Compound AI Systems
Lingjiao Chen (Stanford University), James Zou (Stanford University)
CodeLarge Language ModelTextMultimodalityPhysics Related
π― What it does: This study investigates the non-monotonic characteristics of the performance of composite systems that involve multiple calls to language models, followed by voting or filtered voting, as the number of calls varies across different language tasks. A theoretical explanation centered on query difficulty and a predictable scale model are proposed.
Are Self-Attentions Effective for Time Series Forecasting?
Dongbin Kim (Seoul National University), Hoki Kim (Chung-Ang University)
CodeTransformerTime Series
π― What it does: This paper proposes a time series Transformer called CATS that uses only cross-attention, eliminating self-attention and focusing on future moment predictions.
Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage?
Maohao Shen (Massachusetts Institute of Technology), Gregory W. Wornell (Massachusetts Institute of Technology)
CodeAnomaly DetectionKnowledge DistillationImage
π― What it does: This paper systematically analyzes and reveals the fundamental limitations of existing evidence deep learning (EDL) methods in uncertainty quantification, and proposes an improved scheme based on bootstrap sampling called Bootstrap-Distill.
CodeGenerationRetrievalComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: ARKVALE dynamically identifies and recalls important pages by dividing the KV cache into pages, asynchronously backing them up to external memory, and using bounding volume compressed summaries, supporting long-context LLM inference.
CodeTransformerDiffusion modelAuto EncoderTime SeriesPhysics Related
π― What it does: The AROMA framework is proposed, which compresses spatial information through fixed-size latent markers on arbitrary geometries and allows querying of prediction results at any location.
ART: Automatic Red-teaming for Text-to-Image Models to Protect Benign Users
Guanlin Li (Nanyang Technological University), Tianwei Zhang (Nanyang Technological University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageText
π― What it does: An automatic red team framework named ART is proposed to discover security risks in text-to-image models under the premise of safe text prompts.
Artemis: Towards Referential Understanding in Complex Videos
Jihao Qiu (University of Chinese Academy of Sciences), Yunjie Tian (University at Buffalo)
CodeObject DetectionObject TrackingGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
π― What it does: This paper proposes the Artemis multimodal large language model, specifically designed to locate targets in given frames of videos and generate complete action descriptions.
π― What it does: An unsupervised perspective-synthesis-based neural radiance field method is proposed for learning part segmentation and pose of articulated objects.
Artificial Generational Intelligence: Cultural Accumulation in Reinforcement Learning
Jonathan Cook (University of Oxford), Jakob Nicolaus Foerster
CodeMeta LearningReinforcement LearningSequential
π― What it does: This paper studies two mechanisms for achieving cultural accumulation in reinforcement learning: in-context rapid adaptation and in-weights accumulation, demonstrating that both methods can continuously improve performance across multiple generations of training.
Association Pattern-aware Fusion for Biological Entity Relationship Prediction
Lingxiang Jia (Zhejiang University), Mingli Song (Zhejiang University)
CodeDrug DiscoveryGraph Neural NetworkTransformerBiomedical Data
π― What it does: A method for predicting biological entity relationships based on association pattern-aware fusion, called Pattern-BERP, is proposed.
π― What it does: A distributed acceleration framework named AsyncDiff is proposed, which achieves asynchronous denoising by splitting the denoising network into several sub-modules and executing them in parallel on multiple GPUs, significantly reducing inference latency.
π― What it does: Designed and implemented the ZoDiac watermarking framework, which embeds invisible watermarks in the latent space using a pre-trained Stable Diffusion model, and can still be reliably detected after the images undergo various attacks.
π― What it does: This paper proposes to improve the transfer performance in Cross-Domain Few-Shot Learning (CDFSL) by adjusting the attention temperature (even setting it to 0) in Vision Transformer (ViT), and provides corresponding source domain training and target domain inference strategies.
Attractor Memory for Long-Term Time Series Forecasting: A Chaos Perspective
Jiaxi Hu (Hong Kong University of Science and Technology), Yuxuan Liang
CodeOptimizationTime Series
π― What it does: A long-term time series forecasting model based on chaos theory, Attraos, is proposed, which captures chaotic attractors using phase space reconstruction, non-parametric embedding, and multi-scale dynamic memory units, and performs local evolution in the frequency domain.
Boyu Han (Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: This paper proposes AUCSeg, an AUC optimization method for pixel-level long-tail semantic segmentation, combined with a Tail-Class Memory Bank to address large batch memory issues.
Autoformalize Mathematical Statements by Symbolic Equivalence and Semantic Consistency
Zenan Li (Nanjing University), Xiaoxing Ma (Nanjing University)
CodeTransformerLarge Language ModelText
π― What it does: This paper designs an automatic formalization framework that scores and filters multiple candidates generated by LLMs using two self-consistency methods: symbolic equivalence and semantic consistency, in order to improve the accuracy of automatic formalization of mathematical expressions.
AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning
Minghao Chen (Hangzhou Dianzi University), Xiaofei He (Zhejiang University)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Designed and implemented the AutoManual framework, allowing large language model agents to self-construct environment manuals through planning and rule updates in interactive environments, achieving adaptive learning.
Automated Label Unification for Multi-Dataset Semantic Segmentation with GNNs
Rong Ma (Fudan University), Jian Pu (Fudan University)
CodeObject DetectionSegmentationGraph Neural NetworkLarge Language ModelImage
π― What it does: A unified label space for multiple datasets is automatically constructed using Graph Neural Networks (GNN), enabling the semantic segmentation model to train simultaneously on seven different datasets and output a unified label.
Mengxi Zhang (Baidu Inc.), Yifan Sun (Chinese Academy of Science)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
π― What it does: An Automated Multi-level Preference (AMP) framework has been developed, which significantly reduces hallucinations in multimodal large language models by generating a multi-level preference dataset without human annotations and using the MDPO algorithm.
Automated Multi-Task Learning for Joint Disease Prediction on Electronic Health Records
Suhan Cui (Pennsylvania State University), Prasenjit Mitra (Pennsylvania State University)
CodeOptimizationMeta LearningTabularBiomedical DataElectronic Health Records
π― What it does: AutoDP is proposed, an automated multi-task learning framework for jointly predicting multiple diseases on electronic health records (EHR).
Automatically Learning Hybrid Digital Twins of Dynamical Systems
Samuel Holt (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeLarge Language ModelTime SeriesBiomedical Data
π― What it does: Automatically generate and optimize hybrid digital twins (HDTwin) driven by LLM-based evolutionary algorithms, while integrating mechanistic equations and neural networks to achieve precise modeling of dynamic systems.
Automating Data Annotation under Strategic Human Agents: Risks and Potential Solutions
Tian Xie (Ohio State University), Xueru Zhang (Ohio State University)
CodeOptimizationData-Centric LearningTabularFinance Related
π― What it does: This paper studies the strategic interaction between agents and the long-term dynamics of models when machine learning models are periodically retrained and use model-labeled and human-labeled samples. It theoretically derives the evolution laws of acceptance rate, qualification rate, and classifier bias; further, it proposes a refined retraining strategy using probabilistic samplers to stabilize the system and analyzes its long-term impact on algorithmic fairness.
Autonomous Agents for Collaborative Task under Information Asymmetry
Wei Liu (Tsinghua University), Chen Qian (Tsinghua University)
CodeLarge Language ModelAgentic AITextBenchmark
π― What it does: Proposed the iAgents framework to address collaborative tasks of multi-agent systems under information asymmetry, and constructed the first benchmark for this scenario - InformativeBench.
Jianqiao Lu (University of Hong Kong), Zhijiang Guo (Chinese University of Hong Kong)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The AutoPSV method is proposed, which enhances the reasoning ability of LLMs by training a result supervision validator and automatically generating process labels for each step based on the model's own confidence changes.
π― What it does: A self-regressive image diffusion model (AID) is proposed for generating image sequences and serving as a generative prior in accelerated MRI reconstruction.
π― What it does: A new Diffusion Loss is proposed, allowing autoregressive image generation models to directly use continuous value tokens, eliminating the vector quantization step.
AutoTimes: Autoregressive Time Series Forecasters via Large Language Models
Yong Liu (Tsinghua University), Mingsheng Long (Tsinghua University)
CodeTransformerLarge Language ModelPrompt EngineeringTime Series
π― What it does: AutoTimes is proposed, transforming off-the-shelf decoder-based large language models (LLMs) into autoregressive time series predictors by embedding time series segments in the hidden space of the LLM and generating future time steps using the autoregressive capabilities of the LLM. It supports predictions of arbitrary lengths and introduces position information based on text timestamps and context prompts (in-context forecasting).
π― What it does: This paper proposes the All-in-one Video Restoration Network AverNet, specifically designed for recovering from time-varying unknown degradations (TUD).
AWT: Transferring Vision-Language Models via Augmentation, Weighting, and Transportation
Yuhan Zhu (Nanjing University), Limin Wang (Nanjing University)
CodeClassificationDomain AdaptationTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodality
π― What it does: Proposes the AWT (Augment-Weight-Transport) framework, which enhances the zero/few-shot adaptation performance of pre-trained vision-language models using visual and textual augmentation, entropy weighting, and optimal transport.
π― What it does: Proposes the B-ary Tree Push-Pull (BTPP) distributed learning algorithm, which efficiently propagates parameters and gradients using two tree structures;
BAdam: A Memory Efficient Full Parameter Optimization Method for Large Language Models
Qijun Luo (Chinese University of Hong Kong), Xiao Li (Chinese University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A BAdam method based on the Block Coordinate Descent (BCD) framework and Adam update rule is proposed for full parameter fine-tuning of large language models under memory-constrained conditions.
Balancing Context Length and Mixing Times for Reinforcement Learning at Scale
Matthew Riemer (IBM Research), Sarath Chandar (Mila)
CodeTransformerReinforcement LearningTabular
π― What it does: This study investigates the trade-off between the context length of policies and the mixing time in large, partially observable reinforcement learning environments, providing new theoretical bounds that are subsequently validated through experiments.
BAN: Detecting Backdoors Activated by Adversarial Neuron Noise
Xiaoyun Xu (Radboud University), Stjepan Picek (Radboud University)
CodeAdversarial AttackSupervised Fine-TuningImage
π― What it does: This paper proposes a method using Bait Adversarial Neuron Noise (BAN) to detect and defend against backdoor attacks in deep learning models, combining feature space masking to separate positive and negative features.
Stephen Pasteris (Alan Turing Institute), Mark Herbster (University College London)
CodeRecommendation SystemOptimizationReinforcement Learning from Human FeedbackReinforcement LearningGraph
π― What it does: A new Confidence-Based Abstraction (CBA) algorithm is proposed to address the bandwidth feedback prediction problem with zero-reward abandonment actions.
Batched Energy-Entropy acquisition for Bayesian Optimization
Felix Teufel (Novo Nordisk), Jesper Ferkinghoff-Borg (Novo Nordisk)
CodeOptimizationTabular
π― What it does: A statistical physics-based energy-entropy (BEEBO) acquisition function is proposed for high-dimensional large-batch Bayesian optimization, supporting Gaussian processes and heteroscedastic noise, and can be directly optimized for batch points using gradient methods.
Rafael Oliveira (CSIRO Data61), Edwin V. Bonilla (CSIRO Data61)
CodeOptimizationTabular
π― What it does: BACON is proposed, a Bayesian Adaptive Calibration and Optimal Design method that uses information gain to guide simulation design for efficient calibration of computational models.
π― What it does: A domain index inference algorithm GMDI based on Gaussian Mixture Models is proposed for domain adaptation in the absence of domain index information.
Nicola Bariletto (University of Texas at Austin), Nhat Ho (University of Texas at Austin)
CodeOptimizationTabular
π― What it does: A data-driven distributionally robust optimization criterion is proposed, combining Bayesian nonparametrics (Dirichlet process) with smooth ambiguity aversion theory, and a computable approximate form is provided.
Matt Jones (University of Colorado), Kevin Patrick Murphy
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: An online Bayesian natural gradient (BONG) method is proposed, which updates the expected likelihood using a single-step natural gradient, eliminating KL regularization, and is suitable for sequential inference in neural networks.
Bayesian Optimisation with Unknown Hyperparameters: Regret Bounds Logarithmically Closer to Optimal
Juliusz Ziomek (University of Oxford), Michael A Osborne
CodeOptimizationHyperparameter SearchTabular
π― What it does: To address the challenge of unknown length scales in Bayesian optimization, a length scale balancing algorithm LB-GP-UCB is proposed, improving upon the previous A-GP-UCB and providing a tighter cumulative regret upper bound.
π― What it does: This paper studies the precise inversion problem of diffusion model sampling, proposing a general bidirectional explicit linear multi-step sampling framework (BELM), and based on this, designs an optimal sampler (O-BELM) to achieve training-independent precise inversion and high-quality sampling.
BendVLM: Test-Time Debiasing of Vision-Language Embeddings
Walter Gerych (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
CodeClassificationRecognitionData-Centric LearningTransformerLarge Language ModelVision Language ModelImageMultimodality
π― What it does: This paper proposes a testing-time, non-fine-tuning VLM debiasing method called BEND-VLM, which removes gender/racial bias in visual language models by orthogonalizing the local attribute subspace of each query and equalizing reference images.
π― What it does: An improved version of the multilayer perceptron, RealMLP, and tuned default parameters for GBDT are proposed, and their performance is evaluated on large-scale meta-training/meta-testing benchmarks.
Beyond Accuracy: Ensuring Correct Predictions With Correct Rationales
Tang Li (University of Delaware), Xi Peng (University of Delaware)
CodeRetrievalOptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningImage
π― What it does: A dual-positive prediction framework is proposed, which constructs a reason dataset with a hierarchical structure and designs a reason-based unsupervised optimization method to enhance the model's prediction accuracy and reason interpretability.
π― What it does: A target tracking algorithm called CPDTrack is proposed, which is based on the human visual search mechanism and can continuously locate targets and quickly recover in complex scenes.
Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable?
Sonia Laguna (ETH Zurich), Julia E Vogt
CodeSupervised Fine-TuningVision Language ModelImageTabularBiomedical Data
π― What it does: Implement instance-level concept intervention on existing black-box neural networks and enhance intervention effectiveness through fine-tuning on a small validation set.
Beyond Optimism: Exploration With Partially Observable Rewards
Simone Parisi (University of Alberta), Michael Bowling (University of Alberta)
CodeOptimizationReinforcement LearningTabular
π― What it does: A novel exploration strategy driven by Successor Representation (SR) that does not rely on optimistic assumptions is proposed and its effectiveness is validated within the framework of Monitored MDP with partially observable rewards.
π― What it does: In unsupervised multi-layer graph structure learning, this paper proposes the InfoMGF framework, which first refines the graph structure for each view to remove irrelevant noise, and then merges all views to generate a global graph that contains both shared and unique task-related information for learning node representations.
Beyond Single Stationary Policies: Meta-Task Players as Naturally Superior Collaborators
Haoming Wang (Xi'an Jiaotong University), Zhongmin Cai (University of Notre Dame)
CodeRobotic IntelligenceMeta LearningReinforcement Learning from Human FeedbackReinforcement LearningVideo
π― What it does: A human-machine collaboration framework named Collaborative Bayesian Policy Reuse (CBPR) has been designed and implemented, which can dynamically select the most suitable meta-task-specific strategy based on the non-stationary behavior of human partners, thereby achieving more efficient collaboration.
π― What it does: This paper unifies previous cryptanalysis extraction methods to construct a complete codebase for end-to-end extraction of neural network parameters (weights and symbols), and conducts a systematic evaluation on standard benchmarks (MNIST, CIFAR-10), significantly improving extraction speed and robustness.
π― What it does: The study investigates the sequential multi-task linear bandit problem under the assumption of no task diversity and proposes the BOSS algorithm, which can learn and transfer low-rank representations.
Bias Amplification in Language Model Evolution: An Iterated Learning Perspective
Yi Ren (University of British Columbia), Danica J. Sutherland (University of British Columbia)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This study explores the similarities between large language models (LLMs) and human cultural evolution during the process of self-iterative improvement. It utilizes a Bayesian Iterative Learning (IL) framework to explain the sampling and learning behaviors of LLMs, and experimentally verifies the bias amplification mechanism and the regulatory effects of the interaction phase.
π― What it does: A bidirectional recursive framework GPTrack based on Gaussian process latent encoding is proposed for accurate tracking of cardiac motion.
π― What it does: The BRO (Bigger, Regularized, Optimistic) algorithm is proposed and implemented, significantly improving sample efficiency in continuous control tasks through a large-scale regularized critic network, BroNet structure, optimistic exploration, and non-pessimistic quantized Q approximation.
π― What it does: A binary diffusion model (BI-DiffSR) is proposed for image super-resolution tasks, aiming to significantly reduce memory and computational consumption.
Michael Dinitz (Johns Hopkins University), Sergei Vassilvitskii (Google Research)
CodeTime Series
π― What it does: This study investigates an algorithm that uses distributed predictions in binary search, proposing a robust algorithm that maintains low query complexity even when the prediction distribution is imprecise.
Binding in hippocampal-entorhinal circuits enables compositionality in cognitive maps
Christopher Kymn, Bruno Olshausen
Code
π― What it does: A normative model based on the Residue Number System (RNS) and binding vectors is proposed to explain the role of the hippocampus-olfactory cortex in spatial representation and path integration, further achieving efficient and error-correcting spatial encoding through a modular attractor network.
Biologically Inspired Learning Model for Instructed Vision
Roy Abel (Weizmann Institute of Science), Shimon Ullman (Weizmann Institute of Science)
CodeClassificationRecognitionImage
π― What it does: A biologically interpretable visual learning model that combines the upward (BU) and downward (TD) pathways is proposed, with TD feedback used simultaneously for visual guidance and learning.
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The BISCOPE method is proposed, which captures the forward and backward token information in the logits of LLM outputs through bidirectional cross-entropy loss for detecting AI-generated text.
BitDelta: Your Fine-Tune May Only Be Worth One Bit
James Liu (Massachusetts Institute of Technology), Tianle Cai (Princeton University)
CodeCompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes the BitDelta method, which performs 1-bit binary compression on the weight increments generated by LLM fine-tuning, enabling multiple models to share the same base model.